Incremental semantic segmentation method for deviation context information correction

Abstract

The invention discloses an incremental semantic segmentation method for deviation context information correction. The method comprises the following steps: firstly, acquiring semantic segmentation data streams of a plurality of categories, and dividing the semantic segmentation data streams into a plurality of training data sets; in the first incremental semantic segmentation learning step, an initial semantic segmentation network model is obtained through learning of a training data set; in the next incremental semantic segmentation learning step, a new class pixel point erasing method is used for generating a deviation context information corrected picture pair for a newly obtained training data set containing a new class, constructing a deviation context information corrected training data set, and constructing a deviation context information corrected training data set based on the deviation context information corrected training data set. And updating the latest incremental semantic segmentation network model by a learning method of deviation context information correction and adaptive class balance. According to the method, context information, deviating to a new class, of old-class pixel points can be effectively corrected, the class distribution problem of deviation can be relieved, and forgetting of old-class knowledge and semantic drift of background classes are reduced.

Description

technical field [0001] The invention relates to the field of incremental semantic segmentation, in particular to an incremental semantic segmentation method for correcting biased context information. Background technique [0002] Semantic segmentation is a classic pixel-level classification problem in computer vision. Deep learning methods have yielded amazing results on semantic segmentation tasks when given a large-scale pixel-level labeled dataset. However, in more practical incremental semantic segmentation scenarios, deep neural networks are required to learn a series of tasks with incremental classes and data. The research of incremental semantic segmentation aims to alleviate the network's forgetting of past tasks and over-matching of current tasks without past data. There are two main challenges: first, catastrophic forgetting, in learning the current task When the data that appeared in the past cannot be obtained, the performance of the model in the old task will ...

AI Technical Summary

Problems solved by technology

Compared with the old image, the context of the old class pixels in the new image is more biased towards the new class, which may lead to a sharp deterioration of old class forgetting and new class overmatching

Images